NEUROCOMPUTER FOR IMAGE/SPECTRAL PATTERN RECOGNITION AND ANALYSIS

NEUROPHYSIOLOGICAL RESEARCH IN VISION HAS SHOWN THAT MANY ASPECTS OF PREATTENTIVE VISION PROCESSING ENTAIL SIMPLE ALGORITHMS WHICH ARE IMPLEMENTED IN A HIGHLY PARALLEL ARCHITECTURE. LIKEWISE, ARTIFICIAL NEURAL SYSTEM RESEARCH AND APPLICATIONS HAVE RESULTED IN MANY PARADIGMS WHICH SHARE MANY OF THE SAME PROCESSING STEPS. WE PROPOSE TO DESIGN A SINGLE, UNIFIED TWO-DIMENSIONAL NEURAL NET PROCESSOR ARCHIETECTURE CAPABLE OF DUPLICATING THE MOST IMPORTANT "LOCAL" VISION PROCESSING ALGORITHMS. THE ARCHITECTURE OF THIS PATTERN MATCHING ENGINE WILL ADDRESS THE PROBLEMS OF MEMORY, PROGRAMMING AND PARALLEL IMPLEMENTATION BY EXECUTING A PRIMITIVE INSTRUCTION SET BASED UPON TWO-DIMENSIONAL ADDRESSING. IN PARTICULAR, EMPHASIS WILL BE PLACED ON STEP-AND-REPEAT AND LOCAL AREA TRANSFORMS. THIS PROCESSOR IS TERMED A "COMPUTATIONAL MAP PROCESSOR" OR CMP.